YOLO-Based Depth Estimation from Monocular Camera Images for Object Detection at Sea

It is considered that to be accepted, autonomous ships must have situational awareness that is at least as good as a human look-out. It is important to strive towards the development of trustworthy situational awareness modules for autonomous ships. In this work, we detail our approach to the detection of navigational buoys and monocular depth estimation, one of the challenges of the 3rd Maritime Computer Vision Workshop. The challenge organizers provided an implementation of a YOLOv7 model with a custom head that outputs an additional number representing the metric distance from the camera to the annotated object along with its standard class and bounding box. We found that the precision of the annotations of the objects significantly affected model performance. Although a model trained on our refined labels achieved higher accuracy on the validation set which was annotated by the same person, it performed poorly on the challenge test dataset. Fuhrthermore, we observed that YOLOv7 struggles with detecting small objects, especially at larger distances. To mitigate this, we introduced a random image scaling factor during training, which improved the model's ability to detect smaller objects and increased mAP metrics. Our best-performing model achieved an average distance estimation error of approximately 50 meters across a dataset with object distances ranging from 10 to 1400 meters. Given the challenge of detecting such small objects at extreme distances with YOLOv7, in addition to estimating their metric distance, this result demonstrates the potential of real-time object detection for maritime situational awareness.